[AVX-512] Fix a bug in getLargestLegalSuperClass where we inflated to VR128X/VR256X...

This seems to have been responsible for the XMM16-31 spills observed in PR29112. With this fixed the test case has been modified to no longer have a spill of XMM16.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@283668 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Target/X86/X86RegisterInfo.cpp b/lib/Target/X86/X86RegisterInfo.cpp
index 4401d1769baa2a2b0cc0c483f91a066df2af5d47..87175c32ddce0ba8a9d8aa1aebf1ace654ed440f 100644 (file)
--- a/lib/Target/X86/X86RegisterInfo.cpp
+++ b/lib/Target/X86/X86RegisterInfo.cpp
@@ -146,19 +146,18 @@ X86RegisterInfo::getLargestLegalSuperClass(const TargetRegisterClass *RC,
       if (!Subtarget.hasVLX() && Super->getSize() == RC->getSize())
         return Super;
       break;
-    case X86::FR32XRegClassID:
-    case X86::FR64XRegClassID:
-      // If VLX isn't support we shouldn't inflate to these classes.
-      if (!Subtarget.hasVLX())
-        break;
-      // The VLX check above passed, AVX512 check below will pass.
-      LLVM_FALLTHROUGH;
     case X86::VR128XRegClassID:
     case X86::VR256XRegClassID:
+      // If VLX isn't support we shouldn't inflate to these classes.
+      if (Subtarget.hasVLX() && Super->getSize() == RC->getSize())
+        return Super;
+      break;
+    case X86::FR32XRegClassID:
+    case X86::FR64XRegClassID:
       // If AVX-512 isn't support we shouldn't inflate to these classes.
-      if (!Subtarget.hasAVX512())
-        break;
-      LLVM_FALLTHROUGH;
+      if (Subtarget.hasAVX512() && Super->getSize() == RC->getSize())
+        return Super;
+      break;
     case X86::GR8RegClassID:
     case X86::GR16RegClassID:
     case X86::GR32RegClassID:

diff --git a/test/CodeGen/X86/pr29112.ll b/test/CodeGen/X86/pr29112.ll
index e39759a8c74985326ec7430e9c467895f363178d..08d6aa693156dc1687aed195aaefd049d9900326 100644 (file)
--- a/test/CodeGen/X86/pr29112.ll
+++ b/test/CodeGen/X86/pr29112.ll
@@ -1,11 +1,71 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
 ; RUN: llc < %s -mtriple=x86_64-unknown -mattr=+avx512f | FileCheck %s
 
 declare <4 x float> @foo(<4 x float>, <4 x float>, <4 x float>, <4 x float>, <4 x float>, <4 x float>, <4 x float>, <4 x float>, <4 x float>, <4 x float>)
 
-; In AVX512 without VLX we can't spill XMM16-31 with vmovaps as its not available. Instead we need to use vextractf32x4 to spill since it can encode the ZMM super register and can store the lower 128-bits.
+; Due to a bug in X86RegisterInfo::getLargestLegalSuperClass this test case was trying to use XMM16 and spill it without VLX support for the necessary store instruction. We briefly implemented the spill using VEXTRACTF32X4, but the bug in getLargestLegalSuperClass has now been fixed so we no longer use XMM16.
 
 define <4 x float> @bar(<4 x float>* %a1p, <4 x float>* %a2p, <4 x float> %a3, <4 x float> %a4, <16 x float>%c1, <16 x float>%c2) {
-; CHECK: vextractf32x4 $0, %zmm16, {{[0-9]+}}(%rsp) {{.*#+}} 16-byte Folded Spill
+; CHECK-LABEL: bar:
+; CHECK:       # BB#0:
+; CHECK-NEXT:    subq $88, %rsp
+; CHECK-NEXT:  .Ltmp0:
+; CHECK-NEXT:    .cfi_def_cfa_offset 96
+; CHECK-NEXT:    vmovaps %xmm1, {{[0-9]+}}(%rsp) # 16-byte Spill
+; CHECK-NEXT:    vextractf32x4 $1, %zmm3, %xmm1
+; CHECK-NEXT:    vextractf32x4 $1, %zmm2, %xmm8
+; CHECK-NEXT:    vinsertps {{.*#+}} xmm9 = xmm8[0],xmm1[0],xmm8[2,3]
+; CHECK-NEXT:    vinsertps {{.*#+}} xmm0 = xmm9[0,1],xmm2[1],xmm9[3]
+; CHECK-NEXT:    vinsertps {{.*#+}} xmm15 = xmm0[0,1,2],xmm3[1]
+; CHECK-NEXT:    vblendps {{.*#+}} xmm4 = xmm8[0],xmm1[1],xmm8[2,3]
+; CHECK-NEXT:    vinsertps {{.*#+}} xmm4 = xmm4[0,1],xmm2[1],xmm4[3]
+; CHECK-NEXT:    vinsertps {{.*#+}} xmm6 = xmm4[0,1,2],xmm3[1]
+; CHECK-NEXT:    vmovdqa %xmm6, {{[0-9]+}}(%rsp) # 16-byte Spill
+; CHECK-NEXT:    vextractf32x4 $2, %zmm3, %xmm4
+; CHECK-NEXT:    vblendps {{.*#+}} xmm4 = xmm0[0,1,2],xmm4[3]
+; CHECK-NEXT:    vpermilps {{.*#+}} xmm5 = xmm2[3,1,2,3]
+; CHECK-NEXT:    vinsertps {{.*#+}} xmm5 = xmm5[0],xmm1[0],xmm5[2,3]
+; CHECK-NEXT:    vinsertps {{.*#+}} xmm5 = xmm5[0,1],xmm2[1],xmm5[3]
+; CHECK-NEXT:    vinsertps {{.*#+}} xmm5 = xmm5[0,1,2],xmm3[1]
+; CHECK-NEXT:    vmovshdup {{.*#+}} xmm7 = xmm8[1,1,3,3]
+; CHECK-NEXT:    vinsertps {{.*#+}} xmm7 = xmm7[0],xmm1[0],xmm7[2,3]
+; CHECK-NEXT:    vinsertps {{.*#+}} xmm10 = xmm7[0,1],xmm2[1],xmm7[3]
+; CHECK-NEXT:    vblendps {{.*#+}} xmm7 = xmm10[0,1,2],xmm3[3]
+; CHECK-NEXT:    vblendps {{.*#+}} xmm11 = xmm0[0,1,2],xmm3[3]
+; CHECK-NEXT:    vpermilpd {{.*#+}} xmm12 = xmm3[1,0]
+; CHECK-NEXT:    vpermilpd {{.*#+}} xmm13 = xmm1[1,0]
+; CHECK-NEXT:    vextractf32x4 $3, %zmm3, %xmm0
+; CHECK-NEXT:    vinsertps {{.*#+}} xmm1 = xmm8[0],xmm0[0],xmm8[2,3]
+; CHECK-NEXT:    vinsertps {{.*#+}} xmm1 = xmm1[0,1],xmm2[1],xmm1[3]
+; CHECK-NEXT:    vinsertps {{.*#+}} xmm14 = xmm1[0,1,2],xmm3[1]
+; CHECK-NEXT:    vinsertps {{.*#+}} xmm10 = xmm10[0,1,2],xmm3[1]
+; CHECK-NEXT:    vpermilpd {{.*#+}} xmm0 = xmm0[1,0]
+; CHECK-NEXT:    vinsertps {{.*#+}} xmm0 = xmm8[0],xmm0[0],xmm8[2,3]
+; CHECK-NEXT:    vinsertps {{.*#+}} xmm0 = xmm0[0,1],xmm2[1],xmm0[3]
+; CHECK-NEXT:    vinsertps {{.*#+}} xmm0 = xmm0[0,1,2],xmm13[0]
+; CHECK-NEXT:    vinsertps {{.*#+}} xmm1 = xmm8[0],xmm13[0],xmm8[2,3]
+; CHECK-NEXT:    vinsertps {{.*#+}} xmm1 = xmm1[0,1],xmm2[1],xmm1[3]
+; CHECK-NEXT:    vinsertps {{.*#+}} xmm3 = xmm1[0,1,2],xmm3[1]
+; CHECK-NEXT:    vinsertps {{.*#+}} xmm2 = xmm9[0,1],xmm2[3],xmm9[3]
+; CHECK-NEXT:    vinsertps {{.*#+}} xmm2 = xmm2[0,1,2],xmm12[0]
+; CHECK-NEXT:    vaddps %xmm3, %xmm2, %xmm2
+; CHECK-NEXT:    vmovdqa %xmm15, %xmm1
+; CHECK-NEXT:    vmovdqa %xmm1, {{[0-9]+}}(%rsp) # 16-byte Spill
+; CHECK-NEXT:    vaddps %xmm0, %xmm1, %xmm9
+; CHECK-NEXT:    vaddps %xmm14, %xmm10, %xmm0
+; CHECK-NEXT:    vaddps %xmm1, %xmm1, %xmm8
+; CHECK-NEXT:    vaddps %xmm11, %xmm3, %xmm3
+; CHECK-NEXT:    vaddps %xmm0, %xmm3, %xmm0
+; CHECK-NEXT:    vaddps %xmm0, %xmm1, %xmm0
+; CHECK-NEXT:    vmovaps %xmm8, {{[0-9]+}}(%rsp)
+; CHECK-NEXT:    vmovaps %xmm9, (%rsp)
+; CHECK-NEXT:    vmovaps {{[0-9]+}}(%rsp), %xmm3 # 16-byte Reload
+; CHECK-NEXT:    callq foo
+; CHECK-NEXT:    vmovaps {{[0-9]+}}(%rsp), %xmm1 # 16-byte Reload
+; CHECK-NEXT:    vaddps {{[0-9]+}}(%rsp), %xmm1, %xmm1 # 16-byte Folded Reload
+; CHECK-NEXT:    vaddps %xmm0, %xmm1, %xmm0
+; CHECK-NEXT:    addq $88, %rsp
+; CHECK-NEXT:    retq
   %a1 = shufflevector <16 x float>%c1, <16 x float>%c2, <4 x i32> <i32 4, i32 20, i32 1, i32 17>
 
   %a2 = shufflevector <16 x float>%c1, <16 x float>%c2, <4 x i32> <i32 4, i32 21, i32 1, i32 17>